3.2.1 lead-slice — neutral scoring synthesis

Date: 2026-06-16. Method: after the white-team corroboration (found confirmation-biased) and the red-team refutation, the operator chose a neutral scoring pass. Three scorers — debugger-debbie (severity), planner-priti (tracker/ROI/sequencing), architect-alphonso (architectural blast-radius) — independently scored the same flat candidate set against a neutral rubric, drawing on both white and red evidence, with an explicit de-bias instruction (score naming fairly as one candidate; no predetermined answer). Inputs: scoring-debbie-severity.md, scoring-priti-tracker.md, scoring-alphonso-architecture.md. This is the synthesis.

The candidates (flat, as scored)

# Candidate Priority
#1716 coord topology coherent from mission-create→planning (write-side coord split-brain) P0 launch-blocker
#1832 agent action implement claim succeeds but "no workspace resolved" (read-path) P1 bug
#1827 spec-kitty merge post-merge baseline circular failure (unrecoverable) P1 bug
#1891 agent --json broken P1 bug
#1619/#1666 unify execution-context / execution-state — shippable 3.2.1 slice P0 epics
#1878 write-side slice bounded, characterization-first coord/primary write cut P2 umbrella
Naming routing rider route ~14–26 mission_id[:8]→resolve_mid8 + ratchet tripwire + #1993/#1971/#1888/#1900 mixed

The reframe that resolves the panel (alphonso, verified on checkout)

#1716, the #1619 slice, the #1878 write-side slice, and #1832 are not four candidates — they are one surface at four grains. The write-side coord/topology authority is the root; #1832 is its smallest, safest, verifiable instance; the #1619 builder-hardening is the same surface as an internal-invariant fix; the #1878 slice is the same surface at the entry/durability grain.

This dissolves the apparent debbie (#1716-first) vs priti (#1832-first) disagreement: they are picking different grains of the same lead, not different leads. Verified facts that anchor it:

  • ExecutionContext is mutable (context.py:184) and its builder mutates substrate fields after freezing fragments (resolution.py:793-801) → branch_name ≠ branch_ref.target_branch inside the claimed SSOT. The split-brain lives in the authority, not just in non-adoption.
  • The create-path and re-resolve-path are different resolvers (workflow.py:1336→1357→1372), which is exactly what #1832 reproduces; #1832's own ticket-grounded fix is "consume the same resolved context the claim used to create the workspace — single resolution path."
  • resolve_lanes_dir does not exist yet (#1993 extraction is real); merge.py is a 3.3k-line god-module; only 2 context-fragment reads exist tree-wide; the dashboard uses zero context.

Where the three scorers converge

  1. #1832 is the universally-endorsed entry point. debbie #2-by-severity (verified real, reproduced); priti #1-by-ROI (cheapest real-pain fix, is the G2 "consume-the-SSOT-don't-re-derive" thesis as an actual bug not a metaphor); alphonso "the safest possible first WP, proves single-resolution."
  2. #1716 is the highest-severity / highest-blast-radius keystone. debbie #1; alphonso #1 (paired); priti "most important candidate and the worst first move" (it's the deep/risky #1878 keystone).
  3. The naming rider is unanimously demoted to a cheap, safe, parallel track — not the lead. Scored fairly (priti Σ26, alphonso 22/30), neither inflated nor punished. Its impact is low (tidiness debt), and the "ratchet can never regrow" claim is overstated (it passes today with ~20 live [:8] sites). #1888 is verify-and-close, not a build.
  4. Hard sequencing constraint (priti, echoed by alphonso): #1993 must land with #1832, not alone — extracting resolve_lanes_dir by itself half-strangles the read path into a new shadow path.
  5. Tracker inversion is real, not rhetorical (priti, alphonso-confirmed): the 3.2.x milestone holds only the 5 (now-closed) naming issues; #1716 (P0), #1827/#1891 (P1), #1832 (P1), and #1619/#1666 (P0 epics) are all unmilestoned. This needs fixing regardless of which lead is chosen.

Where they diverge (and the resolution)

  • #1827 — priti: highest user pain (unrecoverable in-tool, needs manual meta.json surgery), her runner-up lead. debbie: may be stale — the claimed "assert-before-write" circularity does not reproduce on current code (merge.py orders record→commit→assert correctly, resume-convergent). alphonso: neutral. → Resolution: re-test #1827 first. If it reproduces, it's a high-value fast-follow; if not, close-with-evidence (claim-exempt). Do not lead with it unconfirmed.
  • #1891 — partially fixed already on this branch (4c492aa85 fixed the map-requirements CommitResult); residual (agent action implement has no --json) is independent DevEx, start anytime. Not a lead.

Synthesized recommendation

3.2.1 LEAD = the write-side / single-resolution authority surface — opened through #1832, driving to #1716. Concretely:

  • WP1 — #1832 + #1993 (must pair): single-resolution read-path fix. Safest first WP, lands a live P1, proves "consume the resolved context, don't re-derive," and supplies the resolve_lanes_dir seam.
  • WP-next — #1716: the write-side coord/topology authority root (P0 launch-blocker, keystone of #1619/#1878). Higher characterization-test cost; sequence right after WP1 establishes the pattern.
  • Optional in-cycle — #1619 builder-hardening slice: un-mutate the ExecutionContext builder, close the internal branch_name ≠ branch_ref.target_branch invariant, write the action-vs-bulk guardrail. Same surface, internal-invariant grain.

Parallel tracks (explicitly NOT the lead):

  • Naming routing rider (#2000/#1971/#1900 + ratchet-as-tripwire + #1888 verify-close) — cheap, safe, low-impact; runs alongside. (#1993 goes with WP1, not here.)
  • #1827 — re-test, then fast-follow or close.
  • #1891 residual — independent DevEx, anytime.

This inverts the prior "naming-first" plan the (biased) corroboration had produced: naming becomes a rider, the write-side/single-resolution surface becomes the headline.

The one trade-off that is the operator's to decide

All three scorers flag the same values call (none claims it's a data verdict):

  • Lead with IMPACT — open the write-side/single-resolution surface (#1832→#1716). Frees P0/P1, attacks the #1619/#1878 spine, but carries higher characterization-test cost and semantics risk. (All three scorers lean here.)
  • Lead with SAFETY — open with the byte-identical naming rider. Lowest risk and quick momentum, but frees no P0/P1 and parks the highest-blast-radius surface plus the unrecoverable bug class for later.

Neutral read: impact, entered through #1832 — the safety of the naming-first plan is available inside the impact plan (because #1832 is itself the safest first WP), so leading with impact does not forfeit safety; leading with naming forfeits impact.

Operator decision (2026-06-16)

The operator chose SAFETY — naming routing rider first, deliberately overriding the panel's lean toward impact. This is recorded as a legitimate values choice, not a data verdict: all three scorers leaned impact, and the panel's neutral read is that leading with naming forfeits the highest-blast-radius surface and frees no P0/P1 this patch. The operator weighs lowest-risk momentum and establishing the ratchet/routing pattern above that — a defensible stabilization-cycle stance.

Therefore 3.2.1 = the naming routing rider (route ~20 mission_id[:8]→resolve_mid8, ratchet extension as tripwire, the resolve_lanes_dir/locate_project_root seams, #1888 verify-close). The write-side/single-resolution surface (#1832→#1716), #1827 (re-test first), the #1619 builder-hardening, and the #1891 residual defer to later 3.2.x patches — impact work follows the safe opener.

Two constraints that survive the values call (carry into the mission slice regardless):

  1. #1993 must NOT land alone. Extracting resolve_lanes_dir by itself half-strangles the read path into a new shadow path (the read-side _lanes_feature_dir twin). In the naming-first plan #1832 is deferred, so either (a) #1993 carries a minimal read-side adoption of the new seam in the same WP, or (b) #1993 defers with #1832 to the write-side patch. Decide at slice time — do not ship #1993 as a bare extraction.
  2. The framing in docs/release-goals/3.2.x.md is corrected, not preserved: naming-first is a chosen low-risk opener, not "the data says naming-first." The confirmation-biased "evidence-grounded continuation" claim is discounted; the dialectic + this neutral panel are recorded honestly.

Follow-up (independent of the lead): the verified tracker inversion is fixed — milestone the unmilestoned P0/P1 work (#1716, #1827, #1832, #1891, #1619/#1666) onto 3.2.x/3.3.x so the burndown reflects the real cycle, not only the naming issues.

Edit this page